Global Localization Using Low-frequency Image-based Descriptor and Range Data-based Validation

In image-based global localization, a robot pose is estimated through image association when the robot revisits a previously visited location on a map. Image association is typically performed using high-level local features such as scale invariant feature transform (SIFT) and speeded up robust feature (SURF). However, these methods suffer from false-positive association and high computational load to reject outliers. In this study, we introduce a novel global localization method based on the proposed low-frequency image-based descriptor (LFID) and laser range data. The image is first processed by reducing the range of luminance in the frequency domain. Visual features are then extracted from the processed image through a kernel window. These visual features are described as binary representation for fast association. Because this binary representation includes a spatial distribution of features, it can minimize false-positive association. Nevertheless, false-positive association could occur when scenes appear to be similar from different viewpoints. To address this problem, this study adopts a laser rangefinder to validate the similarity of the place and reject false-positives from the scene recognition. Experimental results confirm the effectiveness of the proposed scheme in actual indoor environments.